gallium/u_threaded: late alloc cpu_storage

Instead of allocating cpu_storage in threaded_resource_init, defer the
allocation to first use (in tc_buffer_map).
This avoids needless memory allocation if tc_buffer_disable_cpu_storage is
called before tc_buffer_map.

map_buffer_alignment is stored and serves as a "can cpu_storage be used" flag.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15074>

src/gallium/auxiliary/driver_noop/noop_pipe.c

@@ -120,7 +120,7 @@ static struct pipe_resource *noop_resource_create(struct pipe_screen *screen,
       FREE(nresource);
       return NULL;
    }
-   threaded_resource_init(&nresource->b.b, false, 0);
+   threaded_resource_init(&nresource->b.b, false);
    return &nresource->b.b;
 }
 

src/gallium/auxiliary/util/u_threaded_context.c

@@ -671,8 +671,7 @@ tc_is_buffer_busy(struct threaded_context *tc, struct threaded_resource *tbuf,
  * allow_cpu_storage should be false for user memory and imported buffers.
  */
 void
-threaded_resource_init(struct pipe_resource *res, bool allow_cpu_storage,
-                       unsigned map_buffer_alignment)
+threaded_resource_init(struct pipe_resource *res, bool allow_cpu_storage)
 {
    struct threaded_resource *tres = threaded_resource(res);
 
@@ -692,7 +691,9 @@ threaded_resource_init(struct pipe_resource *res, bool allow_cpu_storage,
        /* We need buffer invalidation and buffer busyness tracking for the CPU
         * storage, which aren't supported with pipe_vertex_state. */
        !(res->bind & PIPE_BIND_VERTEX_STATE))
-      tres->cpu_storage = align_malloc(res->width0, map_buffer_alignment);
+      tres->allow_cpu_storage = true;
+   else
+      tres->allow_cpu_storage = false;
 }
 
 void
@@ -2173,10 +2174,14 @@ tc_buffer_map(struct pipe_context *_pipe,
    usage = tc_improve_map_buffer_flags(tc, tres, usage, box->x, box->width);
 
    /* If the CPU storage is enabled, return it directly. */
-   if (tres->cpu_storage && !(usage & TC_TRANSFER_MAP_UPLOAD_CPU_STORAGE)) {
+   if (tres->allow_cpu_storage && !(usage & TC_TRANSFER_MAP_UPLOAD_CPU_STORAGE)) {
       /* We can't let resource_copy_region disable the CPU storage. */
       assert(!(tres->b.flags & PIPE_RESOURCE_FLAG_DONT_MAP_DIRECTLY));
 
+      if (!tres->cpu_storage)
+         tres->cpu_storage = align_malloc(resource->width0, tc->map_buffer_alignment);
+
+      if (tres->cpu_storage) {
          struct threaded_transfer *ttrans = slab_alloc(&tc->pool_transfers);
          ttrans->b.resource = resource;
          ttrans->b.level = 0;
@@ -2191,6 +2196,9 @@ tc_buffer_map(struct pipe_context *_pipe,
          *transfer = &ttrans->b;
 
          return (uint8_t*)tres->cpu_storage + box->x;
+      } else {
+         tres->allow_cpu_storage = false;
+      }
    }
 
    /* Do a staging transfer within the threaded context. The driver should

src/gallium/auxiliary/util/u_threaded_context.h

@@ -346,6 +346,7 @@ struct threaded_resource {
     * pointers. */
    bool is_shared;
    bool is_user_ptr;
+   bool allow_cpu_storage;
 
    /* Unique buffer ID. Drivers must set it to non-zero for buffers and it must
     * be unique. Textures must set 0. Low bits are used as a hash of the ID.
@@ -521,8 +522,7 @@ struct threaded_context {
    struct tc_buffer_list buffer_lists[TC_MAX_BUFFER_LISTS];
 };
 
-void threaded_resource_init(struct pipe_resource *res, bool allow_cpu_storage,
-                            unsigned map_buffer_alignment);
+void threaded_resource_init(struct pipe_resource *res, bool allow_cpu_storage);
 void threaded_resource_deinit(struct pipe_resource *res);
 struct pipe_context *threaded_context_unwrap_sync(struct pipe_context *pipe);
 void tc_driver_internal_flush_notify(struct threaded_context *tc);
@@ -613,6 +613,7 @@ tc_buffer_disable_cpu_storage(struct pipe_resource *buf)
       align_free(tres->cpu_storage);
       tres->cpu_storage = NULL;
    }
+   tres->allow_cpu_storage = false;
 }
 
 static inline void

src/gallium/drivers/crocus/crocus_resource.c

@@ -365,7 +365,7 @@ crocus_alloc_resource(struct pipe_screen *pscreen,
    res->base.b.screen = pscreen;
    res->orig_screen = crocus_pscreen_ref(pscreen);
    pipe_reference_init(&res->base.b.reference, 1);
-   threaded_resource_init(&res->base.b, false, 0);
+   threaded_resource_init(&res->base.b, false);
 
    if (templ->target == PIPE_BUFFER)
       util_range_init(&res->valid_buffer_range);

src/gallium/drivers/d3d12/d3d12_resource.cpp

@@ -297,7 +297,7 @@ convert_planar_resource(struct d3d12_resource *res)
          *plane_res = *res;
          d3d12_bo_reference(plane_res->bo);
          pipe_reference_init(&plane_res->base.b.reference, 1);
-         threaded_resource_init(&plane_res->base.b, false, 0);
+         threaded_resource_init(&plane_res->base.b, false);
       }
 
       plane_res->base.b.next = next;
@@ -357,7 +357,7 @@ d3d12_resource_create(struct pipe_screen *pscreen,
    init_valid_range(res);
    threaded_resource_init(&res->base.b,
       templ->usage == PIPE_USAGE_DEFAULT &&
-      templ->target == PIPE_BUFFER, 64);
+      templ->target == PIPE_BUFFER);
 
    memset(&res->bind_counts, 0, sizeof(d3d12_resource::bind_counts));
 
@@ -556,7 +556,7 @@ d3d12_resource_from_handle(struct pipe_screen *pscreen,
    }
    init_valid_range(res);
 
-   threaded_resource_init(&res->base.b, false, 0);
+   threaded_resource_init(&res->base.b, false);
    convert_planar_resource(res);
 
    return &res->base.b;

src/gallium/drivers/freedreno/freedreno_resource.c

@@ -1109,7 +1109,7 @@ alloc_resource_struct(struct pipe_screen *pscreen,
 
    pipe_reference_init(&rsc->track->reference, 1);
 
-   threaded_resource_init(prsc, false, 0);
+   threaded_resource_init(prsc, false);
 
    if (tmpl->target == PIPE_BUFFER)
       rsc->b.buffer_id_unique = util_idalloc_mt_alloc(&screen->buffer_ids);

src/gallium/drivers/iris/iris_resource.c

@@ -479,7 +479,7 @@ iris_alloc_resource(struct pipe_screen *pscreen,
    res->base.b.screen = pscreen;
    res->orig_screen = iris_pscreen_ref(pscreen);
    pipe_reference_init(&res->base.b.reference, 1);
-   threaded_resource_init(&res->base.b, false, 0);
+   threaded_resource_init(&res->base.b, false);
 
    if (templ->target == PIPE_BUFFER)
       util_range_init(&res->valid_buffer_range);

src/gallium/drivers/r600/r600_buffer_common.c

@@ -584,7 +584,7 @@ r600_alloc_buffer_struct(struct pipe_screen *screen,
 	pipe_reference_init(&rbuffer->b.b.reference, 1);
 	rbuffer->b.b.screen = screen;
 
-	threaded_resource_init(&rbuffer->b.b, false, 0);
+	threaded_resource_init(&rbuffer->b.b, false);
 
 	rbuffer->buf = NULL;
 	rbuffer->bind_history = 0;

src/gallium/drivers/radeonsi/si_buffer.c

@@ -571,7 +571,7 @@ static struct si_resource *si_alloc_buffer_struct(struct pipe_screen *screen,
    pipe_reference_init(&buf->b.b.reference, 1);
    buf->b.b.screen = screen;
 
-   threaded_resource_init(&buf->b.b, allow_cpu_storage, SI_MAP_BUFFER_ALIGNMENT);
+   threaded_resource_init(&buf->b.b, allow_cpu_storage);
 
    buf->buf = NULL;
    buf->bind_history = 0;

src/gallium/drivers/zink/zink_resource.c

@@ -904,7 +904,7 @@ resource_create(struct pipe_screen *pscreen,
 
    res->base.b = *templ;
 
-   threaded_resource_init(&res->base.b, false, 0);
+   threaded_resource_init(&res->base.b, false);
    pipe_reference_init(&res->base.b.reference, 1);
    res->base.b.screen = pscreen;